Hugging Face's logo

Androidonnxfork
/
test

Model card Files Community
5
test
File size: 4,970 Bytes 
8b7c501
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
 
// Copyright 2022 Google LLC
//
// This source code is licensed under the BSD-style license found in the
// LICENSE file in the root directory of this source tree.

$assert BATCH_TILE % 8 == 0
$assert BATCH_TILE >= 8
$ABC = "01234567456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
$assert OP in ["ADD", "DIV", "MAX", "MIN", "MUL", "SUB", "SQRDIFF"]
$assert ACTIVATION in ["LINEAR", "MINMAX"]
#include <assert.h>

#include <immintrin.h>

#include <xnnpack/common.h>
#include <xnnpack/intrinsics-polyfill.h>
#include <xnnpack/vbinary.h>


$_MM256_OP_PS = {
$  "ADD": lambda x, y: "_mm256_add_ps(%s, %s)" % (x, y),
$  "DIV": lambda x, y: "_mm256_div_ps(%s, %s)" % (x, y),
$  "MAX": lambda x, y: "_mm256_max_ps(%s, %s)" % (x, y),
$  "MIN": lambda x, y: "_mm256_min_ps(%s, %s)" % (x, y),
$  "MUL": lambda x, y: "_mm256_mul_ps(%s, %s)" % (x, y),
$  "SUB": lambda x, y: "_mm256_sub_ps(%s, %s)" % (x, y),
$  "SQRDIFF": lambda x, y: "_mm256_sub_ps(%s, %s)" % (x, y),
$}[OP]
$SUFFIX = {"LINEAR": "", "MINMAX": "_minmax"}[ACTIVATION]
$PARAMS = {"LINEAR": "xnn_f16_default_params", "MINMAX": "xnn_f16_minmax_params"}[ACTIVATION]
void xnn_f16_v${OP.lower()}${SUFFIX}_ukernel__f16c_x${BATCH_TILE}(
    size_t batch,
    const void* restrict input_a,
    const void* restrict input_b,
    void* restrict output,
    const union ${PARAMS} params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(batch != 0);
  assert(batch % sizeof(uint16_t) == 0);
  assert(input_a != NULL);
  assert(input_b != NULL);
  assert(output != NULL);

  const uint16_t* a = (const uint16_t*) input_a;
  const uint16_t* b = (const uint16_t*) input_b;
  uint16_t* o = (uint16_t*) output;

  $if ACTIVATION == "MINMAX":
    const __m256 vy_min = _mm256_load_ps(params->avx.min);
    const __m256 vy_max = _mm256_load_ps(params->avx.max);

  $if BATCH_TILE > 8:
    for (; batch >= ${BATCH_TILE} * sizeof(uint16_t); batch -= ${BATCH_TILE} * sizeof(uint16_t)) {
      const __m256 va${ABC[0:8]} = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) a));
      const __m256 vb${ABC[0:8]} = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) b));
      $for N in range(8, BATCH_TILE, 8):
        const __m256 va${ABC[N:N+8]} = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (a + ${N})));
        const __m256 vb${ABC[N:N+8]} = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (b + ${N})));
      a += ${BATCH_TILE};
      b += ${BATCH_TILE};

      $for N in range(0, BATCH_TILE, 8):
        __m256 vy${ABC[N:N+8]} = _mm256_cvtph_ps(_mm256_cvtps_ph(${_MM256_OP_PS("va" + ABC[N:N+8], "vb" + ABC[N:N+8])}, _MM_FROUND_TO_NEAREST_INT));

      $if OP == "SQRDIFF":
        $for N in range(0, BATCH_TILE, 8):
          vy${ABC[N:N+8]} = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vy${ABC[N:N+8]}, vy${ABC[N:N+8]}), _MM_FROUND_TO_NEAREST_INT));

      $if ACTIVATION == "MINMAX":
        $for N in range(0, BATCH_TILE, 8):
          vy${ABC[N:N+8]} = _mm256_max_ps(vy${ABC[N:N+8]}, vy_min);

        $for N in range(0, BATCH_TILE, 8):
          vy${ABC[N:N+8]} = _mm256_min_ps(vy${ABC[N:N+8]}, vy_max);

      _mm_storeu_si128((__m128i*) o, _mm256_cvtps_ph(vy${ABC[0:8]}, _MM_FROUND_TO_NEAREST_INT));
      $for N in range(8, BATCH_TILE, 8):
        _mm_storeu_si128((__m128i*) (o + ${N}), _mm256_cvtps_ph(vy${ABC[N:N+8]}, _MM_FROUND_TO_NEAREST_INT));
      o += ${BATCH_TILE};
    }
  for (; batch >= 8 * sizeof(uint16_t); batch -= 8 * sizeof(uint16_t)) {
    const __m256 va = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) a));
    const __m256 vb = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) b));
    a += 8;
    b += 8;

    __m256 vy = _mm256_cvtph_ps(_mm256_cvtps_ph(${_MM256_OP_PS("va", "vb")}, _MM_FROUND_TO_NEAREST_INT));
    $if OP == "SQRDIFF":
      vy = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vy, vy), _MM_FROUND_TO_NEAREST_INT));

    $if ACTIVATION == "MINMAX":
      vy = _mm256_max_ps(vy, vy_min);
      vy = _mm256_min_ps(vy, vy_max);

    _mm_storeu_si128((__m128i*) o, _mm256_cvtps_ph(vy, _MM_FROUND_TO_NEAREST_INT));
    o += 8;
  }
  if XNN_UNLIKELY(batch != 0) {
    const __m256 va = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) a));
    const __m256 vb = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) b));

    __m256 vy = _mm256_cvtph_ps(_mm256_cvtps_ph(${_MM256_OP_PS("va", "vb")}, _MM_FROUND_TO_NEAREST_INT));
    $if OP == "SQRDIFF":
      vy = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vy, vy), _MM_FROUND_TO_NEAREST_INT));

    $if ACTIVATION == "MINMAX":
      vy = _mm256_max_ps(vy, vy_min);
      vy = _mm256_min_ps(vy, vy_max);

    __m128i vh = _mm256_cvtps_ph(vy, _MM_FROUND_TO_NEAREST_INT);
    if (batch & (4 * sizeof(uint16_t))) {
      _mm_storel_epi64((__m128i*) o, vh);
      vh = _mm_unpackhi_epi64(vh, vh);
      o += 4;
    }
    if (batch & (2 * sizeof(uint16_t))) {
      _mm_storeu_si32(o, vh);
      vh = _mm_srli_epi64(vh, 32);
      o += 2;
    }
    if (batch & (1 * sizeof(uint16_t))) {
      *o = (uint16_t) _mm_extract_epi16(vh, 0);
    }
  }
}